Premature optimisation is the process of making design decisions or modifications based on assumed performance improvement without testing, benchmarking and verifying the change.

"Foreign keys are slow so I won't create them" is in fact an example of premature optimisation. (unless someone has sat down, done the measurements and found that in this particular case the foreign keys are unacceptably slow)

"I'm denormalising for performance" is another common case of premature optimisation, someone has assumes that normalised designs are slow and denormalised without testing that assumption (at least I've never encountered a case where they did test).

Optimising queries or tweaking data models or playing with hints or indexes without any measurements and without any idea of the results of the fiddling is premature optimisation

Creating a good, solid, scalable design is not premature optimisation. It's doing a good #@$@# job.